Interpolation kernel machines belong to the class of interpolating classifiers that interpolate all the training data and thus have zero training error. Recent research shows that they do generalize well and have competitive performance. Several recent works proposed various ways of performance improvement for this decision model. In this work we investigate the generalization of interpolation kernel machines, which has not yet received enough attention. Our work is based on the popular regularization formulation with a penalty to the original loss function in order to constrain the model’s capacity. We concretize six regularization methods. The experimental results clearly demonstrate the potential of generalization for classification performance improvement.